Benchmark
Honest numbers. Guard's detection layers evaluated against 69 labeled prompt-injection samples and 20 benign controls. The seed corpus is checked into the repo at corpus/known_attacks.jsonl; the runner is benchmarks/run.py. Numbers on this page reflect the last committed evaluation.
Precision
—
Recall
—
F1
—
p50 latency
—
What's running. The numbers above reflect Guard's regex + Unicode + source-rules layers only. The DeBERTa classifier and the semantic layer (MiniLM + corpus embeddings) are enabled in the deployed Docker image; numbers will update on next deploy.
Confusion matrix
| Metric | Count |
|---|---|
| loading… | |
Per-attack-type hits
| Attack type | True-positive hits |
|---|---|
| loading… | |
How we compare (claimed)
Third-party numbers are aspirational — not run on the same corpus. Treat as directional. We'll publish a head-to-head once we have API credits on each competitor.
| Product | Delivery | Runtime | Entry price | Source-aware |
|---|---|---|---|---|
| Veil Guard | API | Yes | $49/mo | Yes |
| Lakera Guard | API | Yes | Enterprise | Partial |
| Azure Content Safety | API | Yes | Bundled | No |
| Guardrails AI | Python lib | Self-integrated | Free | No |
| Promptfoo | CI | Pre-deploy only | Free | — |
Results generated: —